부산대학교 도서관

Although deep learning excels in sentence-level relation extraction, document-level extraction poses challenges. To address this, we propose a neural network combining local and global entity representations, sourced directly from sentences. We construct mention-level graphs to capture intra and inter-entity connections. Utilizing advanced encoders like Bi-LSTM and BERT, our model effectively learns word representations from entity-linked text. Experiments on DocRED dataset yield impressive 57.33/59.84 Ign F1/F1 scores, a significant improvement over benchmarks.